class: inverse <style> .remark-slide-content > h3 { text-align: left !important; color: white; }; </style> # Semantics of Shame in Social Media Discussions of Reality TV Fans ### DH2019, Utrecht (NL) July 11, 2019 .pull-left.left[ ### Radim Hladík National Institute of Informatics (Japan) Institute of Philosophy of the Czech Academy of Sciences (Czech Republic) ##### Contact radim.hladik@fulbrightmail.org @hlageek ##### Acknowledgement Grant-in-Aid for JSPS Fellows no. 17F17769. ] .pull-right.left[ ### Markéta Štěchová Charles University (Czech Republic) <br><br><br> ##### Contact stechova@fsv.cuni.cz @marquetstech ##### Acknowledgement Czech Science Foundation no. 17-02521S. ] --- # Reality TV and social stratification .large[ - **“Demotic turn”** means that media content becomes more representative of non-elite members of society (Turner 2010) - Even with increased working class presence, __upper classes remain overrepresented__ (Stiernstedt and Jakobsson 2017) - Class distinctions expressed through __cultural tastes and lifestyle differences__ (Piper 2004; Matheson 2007) or as supposedly __individual moral shortcomings__ (Hirdman 2016) - Working-class participants systematically become an object of __ridicule or controversy__ (Eriksson 2016; De Benedictis et al. 2017) ] --- # _Wife Swap_ .large3[ .pull-left.left[ + British Reality TV show first broadcasted in 2003 (ran for 7 years) + Premise: Two wives exchange their homes for one week and both receive a reward for completion of the required swap + Czech license acquired by a private broadcaster TV Nova in 2005 + 10 seasons of the Czech edition by 2018 ] .pull-right.left[ + criticized for __"middle-class gaze"__ (Lyle 2008) - working class as the abject other - spectacle of poverty + emotions and representend inequalities trigger __"political talk" in online discussion forums__ (Graham 2012) ] ] --- # Researching the targets of "shaming" ## RQ: What are the targets of shaming practices among the fans of the show on Facebook? .large3[ - shame = an individual emotion - shaming = enforcement of norms through the generation of negative collective affect and public identification of a trespasser (i.e. attempt of the collective to induce shame) ] ## Approach .large3[ 1. Data 2. Sentiment classification 3. Word embeddings and similarity vectors 4. Interpret differences ] --- class: center, middle, inverse # Data acquire and clean --- # Data + Facebook Page of the show https://www.facebook.com/vymenamanzelek/ + 5-years worth of “postings” of all available types: posts by page (n=1273), comments (n=26383) and replies (n=28459) from January 2012 to March 2017 + scraped from Facebook Graph API + processed (e.g. lemmatization) by NLP tools for the Czech languge .center[ <img src="img/chars_hists.png" height="300"> ] --- class: center, middle, inverse # Sentiment classification separate comments with negative polarity --- # Sentiment classification task .pull-left.left[ + the original intention was to use existing tools: - __Czech SubLex__ 1.0 <br> Czech subjectivity lexicon of 4626 evaluative items - __Neural Monkey__ <br> End-to-end deep learning model for machine translation, sentiment analysis and automatic image captioning in Czech - failed to provide desired levels of accuracy on our dataset - tested first on a pilot of 150 comments, eventually we manually coded 1300 comments ] .pull-right.left[ Ground truth (n = 1300) | classifier | metric | value | :-- | :-- | --: | | one rule baseline | accuracy | 0.61 | lexicon | accuracy | 0.65 | | kappa | 0.28 | neural monkey | accuracy | 0.57 | | kappa | 0.16 | bespoke | accuracy | 0.73 | | kappa | 0.44 ] --- # Bespoke sentiment classifier .pull-left.left[ + Features - Total words count - Negative words count - Neutral words count - Positive words count - Negations count + N-grams - unigrams - bigrams - LSA transformed to 50 dimensions + Algorithm - Support Vector Machines with Linear Kernel ] .pull-right.left[ Measure|Value :-----:|:-----: Accuracy|0.70 |(69 in cross-validation) Kappa |0.39 Sensitivity|0.51 Specificity|0.87 Positive Prediction Value|0.77 Negative Prediction Value|0.67 ] --- # Bespoke sentiment classifier .pull-left.left[ + Features - Total words count - Negative words count - Neutral words count - Positive words count - Negations count + N-grams - unigrams - bigrams - LSA transformed to 50 dimensions + Algorithm - Support Vector Machines with Linear Kernel ] .pull-right.left[ Measure|Value :-----:|:-----: Accuracy|0.70 |(69 in cross-validation) Kappa |0.39 Sensitivity|0.51 __Specificity__|__0.87__ __Positive Prediction Value__|__0.77__ Negative Prediction Value|0.67 ] --- # Sentiment by the type of post and year .center[ <img src="img/sentiment_type.png" height="270"> ] .center[ <img src="img/sentiment_year.png" height="270"> ] ??? Negative emotion not necessarily sparked by the page owner. --- class: center, middle, inverse # Word embeddings compare biases --- # Training of word embeddings .large3[ Correlations between the compound "shaming" and the top 100 MFW vectors in the __negative__ and __nonnegative__ subcorpora (n = 750) ] .pull-left.center[ <img src="img/shaming_corrs.png" height="430"> ] .pull-right.left[ __"Shaming" vector__ Word|Cosine similarity :-----:|:-----: shame |0.6081729 filth |0.4771095 crap |0.3931279 yuck |0.3869079 ...|... <br> __100 MFW vector__ ] ??? This result means that there is something going in the semantics of plain shame and shaming. --- # Sanity check with fastText .center[ <img src="img/shaming_freqs_corrs.png" height="450"> ] --- class: center, middle, inverse # Interpretation identify and categorize words specific to shaming --- # Preliminary results .center[ <img src="img/comparison.png"> ] --- # Conclusions (also quite preliminary) .large3[ + shame - semantic field of "shame" on the _Wife Swap_ Facebook page is a mixture of poverty, sexuality, race + shaming - combination of the semantic field of "shame" and negative sentiment - selectively draws from the wider field of shame and disproportionately targets issues of hygiene and concomitant voluntary breach of privacy - elements of misogyny + Is the concept of "middle-class gaze"a middle-class critique? ] --- # Future work .large3[ + improve accuracy in the sentiment classification task - feature engineering - hyperparameter tuning - larger training set - label more comments - use pre-existing data from other domains + improve reliability of word embeddings - train more models, e.g. n = 1000, and construct similarity vectors from mean values - bonus: observe in/stablity of cosine similarity values + implement clustering as an input to interpretation (?) ] --- # Your turn. <style> .remark-slide-content > h3 { text-align: left !important; font-size: 25px; color: red; } a{ color: red; text-decoration: none; /* turns off background coloring of links */ } </style> .pull-left.center[ <img src="img/qr.png" height="400"> ] .pull-right.left[ ### Semantics of Shame in Social Media Discussions of Reality TV Fans ##### DH2019, Utrecht (NL) July 11, 2019 ##### Radim Hladík @hlageek ##### Markéta Štěchová @marquetstech ] ## **https://hlageek.github.io/reports/dh2019**